Point-cloud-based 3D object detection for autonomous navigation in unmanned ground vehicles

Abstract

Autonomous navigation for UGVs faces significant challenges in detecting objects accurately in complex environments. Despite advancements in 2D object detection, the absence of robust 3D object detection models leave a critical gap in the accurate identification of objects in real-time UGV applications. In this thesis, we propose a novel approach for 3D object detection in the context of autonomous navigation for Unmanned Ground Vehicles (UGVs). The suggested approach uses a two-stage pipeline. Utilizing the additional depth information from the 3D remote Sensor, 3D proposals are generated from the point cloud data in the initial stage. These proposals act as potential foci for the detection of objects. GLENetVR and SE SSD fusion architecture is used in the second stage to train and detect objects inside the suggested bounding boundaries. The two 3D Networks make it possible to more accurately distinguish between objects and the backdrop because they capture the spatial relationships in the volumetric representations of the point clouds. Combining two Neural Network and CNN models requires combining their feature representations, such as concatenation or element-wise combination, to form a combined feature representation used for object recognition. Through comprehensive testing and evaluation of benchmark datasets, we want to show the effectiveness and efficiency of our suggested strategy in comparison to existing 2D object detection methodologies, which are limited by their reliance on only visual information. Our research lays the door for increased safety and dependability in autonomous navigation systems for UGVs by embracing the promise of cloud point-based 3D object identification. Our proposed model has shown superior performance, achieving high accuracy of surpassing both the SE SSD and GLENetVR models.